GNN signal detection method based on attention mechanism in cellular-removed large-scale MIMO system

Abstract

The invention discloses a GNN signal detection method based on an attention mechanism in a cellular-removed large-scale MIMO system, and belongs to the technical field of wireless communication. Aiming at the problem that the traditional EP algorithm adopts independent Gaussian approximation for multi-user posterior distribution and cannot effectively capture interference association among users, the method is combined with a graph neural network and a 2D MHA module, the 2D MHA module captures the interference relation between users through the time attention sub-block, the frequency attention sub-block captures the correlation between different features, the two outputs are spliced and subjected to linear projection to form enhanced feature representation, and meanwhile the fusion capacity of prior and likelihood information is enhanced through the EP module. According to the method, the low calculation complexity is kept while the detection precision is improved, and the adaptability and robustness of the algorithm in different channel and network environments are remarkably enhanced.

Description

Technical Field

[0001] This invention relates to the field of wireless communication technology, specifically to a GNN signal detection method based on an attention mechanism in a decellularized massive MIMO system. Background Technology

[0002] Future mobile communication networks will not only strive for higher data transmission rates, but also emphasize providing stable and reliable communication services over a wide geographical area. Cell-Free Massive MIMO (CF-mMIMO) systems combine the advantages of distributed antennas and massive MIMO technology, effectively improving network coverage and spectrum efficiency through the collaborative work of multiple access points (APs).

[0003] The CF-mMIMO system adopts a user-centric network architecture, in which multiple access points (APs) are connected to a central processing unit (CPU) to collaboratively provide services to all user equipment (UEs). Thanks to this distributed deployment, users have a high probability of accessing at least one AP regardless of their location, effectively overcoming the problems of weak signal and poor coverage in traditional cellular systems at cell edges, and achieving a more balanced quality of service.

[0004] However, signal detection in CF-mMIMO systems faces three main challenges: First, simultaneous communication by multiple users significantly increases inter-user interference, greatly increasing the difficulty for the receiver to accurately reconstruct the original signal under strong interference. Second, the large-scale distributed deployment of access points (APs) fundamentally changes channel characteristics; compared to traditional centralized antenna systems, the favorable propagation characteristics of CF-mMIMO channels are significantly reduced, further increasing the complexity of signal detection. Third, large-scale antenna systems place higher computational resource demands on nonlinear detection algorithms, making their computational complexity unbearable. Therefore, how to reduce computational overhead while ensuring detection accuracy has become a key technical challenge for improving the reliability of CF-mMIMO systems and achieving scalable deployment. Summary of the Invention

[0005] The purpose of this invention is to provide a GNN signal detection method based on an attention mechanism in decellularized massive MIMO systems. This method addresses the technical problems of traditional signal detection algorithms in decellularized massive MIMO systems, such as their inability to effectively capture interference correlations between users, difficulty in adapting to complex channel environments, and high computational complexity. The goal is to achieve signal detection with strong adaptability, high detection accuracy, and low computational overhead, thereby improving the signal detection performance and overall reliability of the system.

[0006] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0007] A GNN signal detection method based on attention mechanism in decellularized massive MIMO systems includes the following steps:

[0008] Step S1: Based on the decellularized massive MIMO system architecture, initialize system parameters, including the inverse of the signed variance, mean offset, graph neural network node features, maximum number of training rounds, batch training block size, learning rate, region range, access point location, and signal-to-noise ratio range.

[0009] Step S2: Based on the inverse of the symbol variance and the mean offset, calculate the mean and variance of the approximate posterior distribution of the current iteration; and based on this, calculate the posterior mean and posterior variance of the transmitted symbols, and input them into the graph neural network module.

[0010] Step S3: In the graph neural network module, the output message vector is calculated through a multilayer perceptron and a two-dimensional multi-head attention module, which is used for updating the nodes of this layer and passing messages to the next layer, and updating the feature vector of the variable node;

[0011] Step S4: Calculate the posterior probability of the graph neural network estimate obtained through iteration, and input it into the output module to calculate the soft sign estimate and derive the estimate variance;

[0012] Step S5: Calculate the expected propagation algorithm factor by using the approximate distribution covariance matrix and the posterior covariance matrix, and pass it into the initialization module in the next iteration. Update the inverse of the sign variance and the mean offset based on the factor.

[0013] Step S6: After multiple iterations, the final symbol detection result is obtained.

[0014] Furthermore, the implementation process of step S2 includes: based on the Gaussian posterior function p (t) Construct the likelihood function p for (x|y) (t) (y|x):

[0015]

[0016] Where, p (t) (x|y) represents the approximate posterior distribution of the t-th iteration; Let K be a K-dimensional posterior mean vector, where K is the number of users; It is a K×K diagonal matrix, and its k-th diagonal element satisfy:

[0017]

[0018] in, It represents the prior accuracy of the k-th variable in the (t-1)-th iteration; It is a vector μ (t) The k-th element represents the posterior mean estimate of the k-th variable in the t-th iteration; It is a matrix Σ (t) The k-th diagonal element represents the variance of the k-th variable in the posterior distribution during the t-th iteration.

[0019] Further, in step S2, the graph neural network module includes multiple variable nodes and factor nodes, wherein each variable node corresponds to a user's transmitted symbol, any two variable nodes are connected through a factor node, and the posterior mean and posterior variance are concatenated to form the attribute vector of the variable node.

[0020] Furthermore, in step S3, the two-dimensional multi-head attention module includes a temporal attention sub-block and a frequency attention sub-block;

[0021] The temporal attention sub-block is used to capture the interference relationship between different users; the frequency attention sub-block is used to capture the correlation between different feature dimensions; the outputs of the two sub-blocks are concatenated and linearly projected to form an enhanced feature representation.

[0022] Furthermore, in step S3, a multilayer perceptron is deployed at each factor node to process the joint feature vector of the input; and a gated recursive unit is deployed at each variable node to update the feature vector of the variable node based on the cascaded aggregation message.

[0023] Furthermore, in step S3, the calculation process of the multilayer perceptron and the two-dimensional multi-head attention module, as well as the update process of the variable node feature vectors, are as follows:

[0024] S31. Construct the joint feature vector of the multilayer perceptron input.

[0025]

[0026] in, a||b represents concatenating vectors a and b along the column direction to form a new column vector; for a fixed receiver k, iterate through all j = 1,...,K, j≠k to calculate.

[0027] S32, Multilayer Perceptron Message Vector Output:

[0028] At each factor node (k,j), The input to the multilayer perceptron of this factor node outputs a message vector. Where M(·) represents the input-output conversion function of the multilayer perceptron. d′ f =N uThe feature dimension of the message vector;

[0029] S33. Construction of a two-dimensional multi-head attention module:

[0030] For each user k, all users from different j By concatenating the data in the row direction, we obtain the input Z for the two-dimensional multi-head attention:

[0031]

[0032] in, R = K-1 is the number of neighboring nodes;

[0033] The time attention sub-block takes Z as input and outputs A. time =MHA(Z), where, MHA(·) represents the attention input-output transformation function;

[0034] Frequency attention sub-blocks in Z T Given input, output A freq =MHA(Z) T ),in,

[0035] A time With A T freq By piecing them together, we obtain Then through learnable matrices Linear projection yields the final output of the two-dimensional multi-head attention module. in,

[0036] S34. Averaging of message vectors and updating of variable node features:

[0037] right Perform averaging to obtain the message vector.

[0038]

[0039] in, express The r-th row; finally, the feature vector of the variable node is updated through this message vector.

[0040] Furthermore, in step S4, the calculation of the soft symbol estimate is obtained by taking the mathematical expectation of the posterior probability distribution estimated by the graph neural network.

[0041] Furthermore, the specific implementation process in step S4 includes:

[0042] In the output module, the soft sign estimate is first calculated based on the softmax probability distribution, and then the variance estimate is derived using the second-order statistic of this distribution.

[0043] Soft symbol estimation satisfy:

[0044]

[0045] Where Ω represents the set of constellation points for the transmitted symbol. is the posterior probability of the k-th user symbol being a in the output of the graph neural network in the t-th iteration, where a is the specific value of the constellation point;

[0046] Estimating variance satisfy:

[0047]

[0048] in, Estimation of symbol a and soft symbol The squared error is obtained by weighting and summing the squared errors of all constellation points with their corresponding posterior probabilities, and then obtaining the estimated variance.

[0049] Furthermore, the implementation process of step S5 includes: when the current iteration number has not reached the maximum iteration number T, updating the inverse of the sign variance and the offset of the sign mean using new statistical information, specifically:

[0050] inverse of sign variance λ (t) satisfy:

[0051]

[0052] The offset γ of the sign variance (t) satisfy:

[0053]

[0054] in, Let K be the K-dimensional posterior mean vector. The symbol estimation vector for the current iteration; Let K be a K×K diagonal matrix, and let the k-th diagonal element be... V (t) Let K be a K×K diagonal matrix, and let the k-th diagonal element be... like Then let At the same time

[0055] Furthermore, in step S6, the final symbol detection result is obtained by selecting the constellation point in the constellation point set that has the smallest Euclidean distance to the soft symbol estimation as the corresponding user's transmitted symbol.

[0056] The beneficial effects of the above scheme are as follows:

[0057] 1. This invention introduces a graph neural network to model the system topology, enabling explicit capture of complex interference relationships between users. This effectively overcomes the performance bottleneck caused by the independent Gaussian approximation in traditional expectation propagation algorithms. Unlike traditional methods that treat users as isolated individuals, this scheme allows the detector to perform collaborative reasoning as if understanding the context of a dialogue, thus significantly improving signal detection accuracy in multi-user interference environments and achieving a lower bit error rate.

[0058] 2. This invention integrates a two-dimensional multi-head attention mechanism during message passing, which dynamically assigns appropriate attention weights to different users and different feature dimensions. This design allows the detector to intelligently focus on key interference information and effective feature representations, rather than performing indiscriminate computation. Therefore, this scheme successfully controls computational complexity within an acceptable range while leveraging graph neural networks to achieve high performance gains, ensuring the algorithm's deployability in practical systems.

[0059] 3. This invention deeply integrates a data-driven graph neural network module with a model-driven expectation propagation algorithm framework to construct a detection platform that combines high robustness and strong adaptability. This scheme does not rely on strict assumptions about the channel model and can adaptively respond to different user distributions, access point deployments, and changing channel conditions through learning. This superior generalization ability makes it flexibly applicable to a variety of complex communication scenarios, providing reliable technical support for building next-generation intelligent wireless communication systems. Attached Figure Description

[0060] Figure 1 This is an iterative flowchart of the detector of the present invention;

[0061] Figure 2 This is a model diagram of the decellularized massive MIMO system of the present invention;

[0062] Figure 3 This is a flowchart of the message passing process in the GNN module of this invention;

[0063] Figure 4 This is a flowchart of the MLP and 2D MHA parameter transfer process of this invention;

[0064] Figure 5 When L=16, N=5, At that time, a comparison chart of the detection performance of the present invention with other traditional algorithms was displayed;

[0065] Figure 6 When L=20, N=5, The graph shows a comparison of the detection performance of this invention with other traditional algorithms. Detailed Implementation

[0066] The present invention will now be described in further detail with reference to the accompanying drawings and specific embodiments.

[0067] It should be noted that, unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application pertains.

[0068] A GNN signal detection method based on attention mechanism in decellularized massive MIMO systems includes the following steps:

[0069] Step S1: Based on the decellularized massive MIMO system architecture, initialize system parameters, including the inverse of the signed variance, mean offset, graph neural network node features, maximum number of training rounds, batch training block size, learning rate, region range, access point location, and signal-to-noise ratio range.

[0070] Step S2: Based on the inverse of the symbol variance and the mean offset, calculate the mean and variance of the approximate posterior distribution of the current iteration; and based on this, calculate the posterior mean and posterior variance of the transmitted symbols, and input them into the graph neural network module.

[0071] Step S3: In the graph neural network module, the output message vector is calculated through a multilayer perceptron and a two-dimensional multi-head attention module, which is used for updating the nodes of this layer and passing messages to the next layer, and updating the feature vector of the variable node;

[0072] Step S4: Calculate the posterior probability of the graph neural network estimate obtained through iteration, and input it into the output module to calculate the soft sign estimate and derive the estimate variance;

[0073] Step S5: Calculate the expected propagation algorithm factor by using the approximate distribution covariance matrix and the posterior covariance matrix, and pass it into the initialization module in the next iteration. Update the inverse of the sign variance and the mean offset based on the factor.

[0074] Step S6: After multiple iterations, the final symbol detection result is obtained.

[0075] The implementation process of each step of the present invention will be described in detail below.

[0076] For a CF-mMIMO system, such as Figure 2 As shown, the system includes There are one single-antenna UE and L access points (APs). The APs and UEs are arbitrarily arranged within a designated area. Each AP is equipped with N antennas. All APs are connected to the CPU, and the CPU acquires the received signals from all APs. The [missing information] is... The channel between the UE and the l-th AP is:

[0077]

[0078] in, 1≤l≤L.

[0079] In one time slot of the uplink, each UE Send its uplink data Its average power is in for -QAM Constellation Collection, s m It is one of the constellation points.

[0080] Step S1: Initialize system parameters

[0081] Figure 1 The diagram shows the iterative flowchart of the detector of this invention, which includes an initialization module, a GNN module, and an output module. During detection, the following parameters need to be passed to the detector: channel matrix H, received signal y, and additive Gaussian noise variance σ. 2 Average energy of transmitted symbols The number of internal message passing iterations L in GNN and the total number of detection iterations T.

[0082] Initialization parameters: t=1, Where, λ (0) I represents the inverse of the sign variance. K Represents a K×K identity matrix, γ (0) The initial offset represents the symbol mean, and t represents the detector iteration number index. This represents the initial hidden state of the k-th variable node in the GNN. It also initializes the maximum number of training rounds, batch size, learning rate, region range, access point location, and signal-to-noise ratio range.

[0083] Through the above initialization and module collaboration, the detector can complete signal detection within T iterations. In each iteration, the GNN module will perform L message passing to optimize feature representation, and finally the output module will output the symbol detection result.

[0084] Step S2: Calculate the posterior mean and variance

[0085] This step constructs a likelihood model based on the Gaussian posterior function, and obtains the posterior statistical features of the transmitted symbols through calculation, providing input for the GNN module.

[0086] Based on the Gaussian posterior function relationship, the approximate posterior distribution of the t-th iteration follows a Gaussian distribution, and the likelihood function p (t) (y|x) satisfies:

[0087] Where, p (t) (x|y) represents the approximate posterior distribution of the t-th iteration; Let K be a K-dimensional posterior mean vector, where K is the number of users; It is a K×K diagonal matrix.

[0088] All APs receive uplink data sent by the UE, and let the received signal of the l-th AP be... All APs receive signals The data is sent to the CPU via the fronthaul link. The CPU combines the received signals from the L APs to obtain the received data.

[0089] At the same time, the CPU acquires the channel estimate. This matrix represents the estimated channel between all UEs and all APs; its first... The vector formed by the (l-1)N+1 to lNth elements of the column is the channel. The estimated value. The transmitted signals of all UEs are...

[0090] To accommodate subsequent real-value signal processing, the CPU will receive complex data. and channel estimates Convert to real number:

[0091]

[0092] in, and These represent the real and imaginary parts of a complex number, respectively. The set of constellation symbols for real numbers corresponding to real values ​​is:

[0093] The detection network is deployed on the CPU. The detector is equipped with a GNN network containing K variable nodes, where the k-th variable node corresponds to the data x of user k. k Any two variable nodes are connected through a factor node, and the (k,j)th factor node is connected to the variable node x. k and x j (k≠j).

[0094] At each factor node, a Multi-Layer Perceptron (MLP) and a 2-Dimensional Multi-Head Attention (2D MHA) module are deployed. The MLP contains modules of size [missing information]. and Two hidden layers and a size of N uThe output layer uses a rectified linear unit (ReLU) as a nonlinear activation at the output of each hidden layer. The input dimension of the MLP is 2N. u +2, output dimension is d′ f 2D MHA consists of two parallel attention branches: a temporal attention sub-block and a frequency attention sub-block, both with an input dimension of R×d′. f .

[0095] Gated Recurrent Units (GRUs) are deployed at variable nodes, with an input dimension of N. u +2, output dimension is

[0096] The detector begins iteration, calculating the posterior mean and posterior variance of the t-th iteration:

[0097]

[0098] in,

[0099] Calculate the variable node x based on the inverse of the signed variance and the posterior mean. k Confidence level:

[0100]

[0101] in, It is a matrix Σ (t) The kth diagonal element, It is a vector λ (t-1) The k-th element, and further calculate x. k Approximate value:

[0102]

[0103] in, It is a vector μ (t) The kth element, It is a vector γ (t-1) The kth element.

[0104] Will and Concatenation, used as variable node x in GNN k Attribute vectors:

[0105]

[0106] Step S3: Message passing and feature update in the GNN module

[0107] This step, using a GNN network structure and leveraging MLP and 2D MHA modules to achieve message exchange and feature optimization between nodes, is the core processing step for signal detection. The specific process is as follows:

[0108] If t = 1, then the eigenvectors of the K variable nodes are initialized, where the initial value of the eigenvector of the k-th variable node is... for:

[0109]

[0110] in, It is a learnable matrix. It is a learnable vector, N u h is the dimension of the node feature vector. k Channel matrix The k-th column vector. If t≠1, proceed to the next step.

[0111] S31: GNN message passing iterative initialization and factor node feature vectors:

[0112] set up This represents the message passing iteration index in the GNN. Calculate the feature vectors of all factor nodes, where the feature vector of the (k,j)th factor node is... In GNNs, iterative message passing (MP) occurs between factor nodes and variable nodes. The message passing process is as follows: Figure 3 As shown. In the factor-to-variable message passing phase, at the... In round-robin message passing, the MLP input vector of the factor node (k,j) for:

[0113]

[0114] in, a||b represents concatenating vectors a and b along the column direction to form a new column vector; for a fixed receiver k, iterate through all j = 1,...,K, j≠k to calculate.

[0115] S32, Message vector output of factor node MLP:

[0116] At each factor node (k,j), the input vector Input the MLP to the factor node and obtain its corresponding output. Where M(·) represents the input-output transformation function of the MLP, d′ f d′ represents the feature dimension of the message vector. f=N u .

[0117] Feature enhancement and message vector generation of S33 and 2D MHA modules:

[0118] For each user k, the message vectors from all different j are... Concatenate along the row direction to obtain the input Z of 2DMHA:

[0119]

[0120] in, R = K-1 represents the number of neighboring nodes.

[0121] Temporal attention sub-block: Input Z, output A time =MHA(Z), where, MHA(·) represents the attention input-output transformation function;

[0122] Frequency attention sub-block: with Z T Given input, output A freq =MHA(Z) T ),in,

[0123] A time With A T freq By piecing them together, we obtain Then through learnable matrices Linear projection yields the final output of the two-dimensional multi-head attention module. in,

[0124] S34. Averaging of message vectors and updating of variable node features:

[0125] right Perform averaging to obtain the message vector.

[0126]

[0127] in express The r-th line in the code. The specific parameter transfer process between the MLP and 2D MHA is as follows: Figure 4 As shown.

[0128] In the message passing phase from variable to factor, cascaded vectors Used to update the variable node at the . wheel feature vector The calculation is as follows:

[0129]

[0130] Where U(·) is a gated recursive unit (GRU) function; Indicates the first Round variable node x k eigenvectors, For the first Round variable node x k eigenvectors; This refers to aggregated messages from neighboring factor nodes; It is a learnable matrix. It is a learnable vector.

[0131] Step S4: Soft Sign Estimation and Variance Derivation

[0132] This step converts the GNN-optimized features into signal estimation results and corresponding uncertainty indices through the output module.

[0133] make if If the value is less than or equal to L, proceed to step S31; otherwise, the iteration ends.

[0134] After completing L rounds of message passing, output the posterior probability estimated based on the GNN model:

[0135]

[0136] Where f(·) is the MLP mapping function, which has a size of and The two hidden layers have an output dimension of |Ω|, which are used to map the node features to a |Ω|-dimensional vector, with each dimension corresponding to the posterior probability of a constellation point. It is the feature vector of the k-th variable node in the L-th round of the GNN.

[0137] Then, the posterior probability is calculated:

[0138]

[0139] in, This represents the posterior probability of the k-th symbol taking the value a, estimated based on the GNN model.

[0140] Let exp(·) be the posterior probability of the k-th user with symbol a estimated by GNN at the t-th iteration. exp(·) maps these unnormalized scores to the positive domain, and finally lets Complete the calculation of the posterior probability of the GNN in the current iteration.

[0141] In the output module, the soft sign estimate is first calculated based on the posterior probability, and then the variance estimate is derived using the second-order statistic of the distribution.

[0142] Soft symbol estimation satisfy:

[0143]

[0144] Estimating variance satisfy:

[0145]

[0146] in, Estimation of symbol a and soft symbol The squared error is obtained by weighting and summing the squared errors of all constellation points with their corresponding posterior probabilities, and then obtaining the estimated variance.

[0147] Step S5: Update the inverse of the signed variance and the offset of the signed mean:

[0148] This step is based on the expectation propagation (EP) algorithm, which uses the detection results of the current iteration to update the prior parameters and provide optimized initial conditions for the next iteration.

[0149] When the current iteration count has not reached the maximum iteration count T, the inverse of the signed variance and the offset of the signed mean are updated using the new statistical information, specifically:

[0150] inverse of sign variance λ (t) satisfy:

[0151]

[0152] The offset γ of the sign variance (t) satisfy:

[0153]

[0154] in, Let K be the K-dimensional posterior mean vector. The symbol estimation vector for the current iteration; Let K be a K×K diagonal matrix, and let the k-th diagonal element be... V (t) Let K be a K×K diagonal matrix, and let the k-th diagonal element be... like Then let At the same time

[0155] Step S6: Iteration termination and final detection result output.

[0156] This step determines the iteration termination condition and outputs the final signal detection result.

[0157] Let t = t + 1. If t is less than or equal to T, then jump to step S2; otherwise, the iteration ends and the result is obtained. real part and the imaginary part And form a vector right Make a firm judgment and use the judgment result as the final test result.

[0158] During the model training phase, random channel matrices, transmitted signals, and Gaussian noise samples are generated based on the corresponding non-cellular network channel model, symbol modulation method, and noise variance.

[0159] Let H (q) x (q) n (q) Let be the channel matrix, transmitted signal, and noise generated randomly for the qth time, respectively, and use the formula y q =H q x q +n q Generate received signal sample y q , where 1≤q≤Q, and Q is the number of training samples.

[0160] Will The GNN-MHA-EP detectors corresponding to steps S3 to S5 are trained using this data, and y q and H (q) As input data, x (q) For the labels, optimize the network parameters of the detector such as MLP, 2D MHA, and GRU until the model loss converges or the maximum number of training rounds is reached.

[0161] During the online detection phase, the parameters H, y, σ of the actual scene are input into the detector. 2 And initialize parameters: inverse of the sign variance. offset γ of the sign mean (0) =0, initial features of GNN variable nodes The iterative process begins, completing multiple rounds of signal detection and parameter updates, and finally outputting the detection results of the user-sent symbols.

[0162] Figure 5 and Figure 6All graphs compare the symbol error rate (SER) performance of different signal detection algorithms in decellularized massive MIMO systems, with the horizontal axis representing the signal-to-noise ratio (SNR). The core objective of combining confidence propagation (BP), minimum mean square error (MMSE), expectation propagation (EP), graph neural network combined with expectation propagation (GNN-EP), and the attention-based GNN combined with expectation propagation algorithm proposed in this invention (GNN-MHA-EP) is to verify the detection performance advantages of the algorithm proposed in this invention.

[0163] Figure 5 The experimental parameters are configured as follows: number of users The number of access points (APs) is L = 16, and the number of antennas per AP is N = 5. As shown in the figure, the symbol error rate of all algorithms decreases with increasing signal-to-noise ratio (SNR), and the GNN-MHA-EP algorithm of this invention exhibits the best performance across the entire SNR range, demonstrating that this invention can effectively improve signal detection accuracy.

[0164] Appendix Figure 6 The experimental parameters are configured as follows: number of users With L = 20 access points (APs) and N = 5 antennas per AP, the system complexity is higher and inter-user interference is more significant. However, it can still be observed from the figure that the GNN-MHA-EP algorithm maintains optimal performance, proving that the algorithm of this invention still has stable high performance and stronger adaptability when the system scale changes.

[0165] As can be seen from the above embodiments, the present invention deeply integrates GNN with attention mechanism and EP algorithm, which not only utilizes the modeling ability of GNN for node association and the focusing ability of attention mechanism for key information, but also improves the detection accuracy through iterative optimization of EP algorithm. It effectively solves the problems of high complexity and poor generalization of traditional detection methods, and can be widely applied to various decellularized large-scale MIMO communication scenarios.

[0166] Finally, it should be noted that any parts of this invention not described in detail are prior art. Those skilled in the art will understand that the above descriptions are merely preferred embodiments of the invention and are not intended to limit the invention. Although the invention has been described in detail with reference to the foregoing examples, those skilled in the art can still modify the technical solutions described in the foregoing examples or make equivalent substitutions for some of the technical features. All modifications and equivalent substitutions made within the spirit and principles of the invention should be included within the scope of protection of the invention.

Claims

1. A GNN signal detection method based on attention mechanism in a decellularized massive MIMO system, characterized in that, Includes the following steps: Step S1: Based on the decellularized massive MIMO system architecture, initialize system parameters, including the inverse of the signed variance, mean offset, graph neural network node features, maximum number of training rounds, batch training block size, learning rate, region range, access point location, and signal-to-noise ratio range. Step S2: Based on the inverse of the symbol variance and the mean offset, calculate the mean and variance of the approximate posterior distribution of the current iteration; and based on this, calculate the posterior mean and posterior variance of the transmitted symbols, and input them into the graph neural network module. Step S3: In the graph neural network module, the output message vector is calculated through a multilayer perceptron and a two-dimensional multi-head attention module, which is used for updating the nodes of this layer and passing messages to the next layer, and updating the feature vector of the variable node; Step S4: Calculate the posterior probability of the graph neural network estimate obtained through iteration, and input it into the output module to calculate the soft sign estimate and derive the estimate variance; Step S5: Calculate the expected propagation algorithm factor by using the approximate distribution covariance matrix and the posterior covariance matrix, and pass it into the initialization module in the next iteration. Update the inverse of the sign variance and the mean offset based on the factor. Step S6: After multiple iterations, the final symbol detection result is obtained.

2. The GNN signal detection method based on attention mechanism in a decellularized massive MIMO system according to claim 1, characterized in that, The implementation process of step S2 includes: based on the Gaussian posterior function p (t) Construct the likelihood function p for (x|y) (t) (y|x): Where, p (t) (x|y) represents the approximate posterior distribution of the t-th iteration; Let K be a K-dimensional posterior mean vector, where K is the number of users; It is a K×K diagonal matrix, and its k-th diagonal element satisfy: in, It represents the prior accuracy of the k-th variable in the (t-1)-th iteration; It is a vector μ (t) The k-th element represents the posterior mean estimate of the k-th variable in the t-th iteration; It is a matrix Σ (t) The k-th diagonal element represents the variance of the k-th variable in the posterior distribution during the t-th iteration.

3. The GNN signal detection method based on attention mechanism in a decellularized massive MIMO system according to claim 2, characterized in that, In step S2, the graph neural network module contains multiple variable nodes and factor nodes, wherein each variable node corresponds to a user's transmitted symbol, any two variable nodes are connected through a factor node, and the posterior mean and posterior variance are concatenated to form the attribute vector of the variable node.

4. The GNN signal detection method based on attention mechanism in a decellularized massive MIMO system according to claim 1, characterized in that, In step S3, the two-dimensional multi-head attention module includes a temporal attention sub-block and a frequency attention sub-block; The temporal attention sub-block is used to capture the interference relationship between different users; the frequency attention sub-block is used to capture the correlation between different feature dimensions; the outputs of the two sub-blocks are concatenated and linearly projected to form an enhanced feature representation.

5. The GNN signal detection method based on attention mechanism in a decellularized massive MIMO system according to claim 1, characterized in that, In step S3, a multilayer perceptron is deployed at each factor node to process the joint feature vector of the input; a gated recursive unit is deployed at each variable node to update the feature vector of the variable node based on the cascaded aggregation message.

6. The GNN signal detection method based on attention mechanism in a decellularized massive MIMO system according to claim 1, characterized in that, In step S3, the calculation process of the multilayer perceptron and the two-dimensional multi-head attention module, as well as the update process of the variable node feature vectors, are as follows: S31. Construct the joint feature vector of the multilayer perceptron input. in, a||b represents concatenating vectors a and b along the column direction to form a new column vector; for a fixed receiver k, iterate through all j = 1,...,K, j≠k to calculate. S32, Multilayer Perceptron Message Vector Output: At each factor node (k,j), The input to the multilayer perceptron of this factor node outputs a message vector. Where M(·) represents the input-output conversion function of the multilayer perceptron. d′ f =N u The feature dimension of the message vector; S33. Construction of a two-dimensional multi-head attention module: For each user k, all users from different j By concatenating the data in the row direction, we obtain the input Z for the two-dimensional multi-head attention: in, R = K-1 is the number of neighboring nodes; The time attention sub-block takes Z as input and outputs A. time =MHA(Z), where, MHA(·) represents the attention input-output transformation function; Frequency attention sub-blocks in Z T Given input, output A freq =MHA(Z) T ),in, A time With A T freq By piecing them together, we obtain Then through learnable matrices Linear projection yields the final output of the two-dimensional multi-head attention module. in, S34. Averaging of message vectors and updating of variable node features: right Perform averaging to obtain the message vector. in, express The r-th row; finally, the feature vector of the variable node is updated through this message vector.

7. The GNN signal detection method based on attention mechanism in a decellularized massive MIMO system according to claim 1, characterized in that, In step S4, the calculation of the soft symbol estimation is obtained by taking the mathematical expectation of the posterior probability distribution estimated by the graph neural network.

8. The GNN signal detection method based on attention mechanism in a decellularized massive MIMO system according to claim 1, characterized in that, The specific implementation process in step S4 includes: In the output module, the soft sign estimate is first calculated based on the softmax probability distribution, and then the variance estimate is derived using the second-order statistic of this distribution. Soft symbol estimation satisfy: Where Ω represents the set of constellation points for the transmitted symbol. is the posterior probability of the k-th user symbol being a in the output of the graph neural network in the t-th iteration, where a is the specific value of the constellation point; Estimating variance satisfy: in, Estimation of symbol a and soft symbol The squared error is obtained by weighting and summing the squared errors of all constellation points with their corresponding posterior probabilities, and then obtaining the estimated variance.

9. The GNN signal detection method based on attention mechanism in a decellularized massive MIMO system according to claim 1, characterized in that, The implementation process of step S5 includes: when the current iteration number has not reached the maximum iteration number T, updating the inverse of the sign variance and the sign mean offset using new statistical information, specifically: inverse of sign variance λ (t) satisfy: The offset γ of the sign variance (t) satisfy: in, Let K be the K-dimensional posterior mean vector. The symbol estimation vector for the current iteration; Let K be a K×K diagonal matrix, and let the k-th diagonal element be... V (t) Let K be a K×K diagonal matrix, and let the k-th diagonal element be... like Then let At the same time 10. The GNN signal detection method based on attention mechanism in a decellularized massive MIMO system according to claim 1, characterized in that, In step S6, the final symbol detection result is obtained by selecting the constellation point in the constellation point set that has the smallest Euclidean distance to the soft symbol estimation as the transmitted symbol for the corresponding user.

Images